1

Title:

Sleep-related spike bursts in HVC are driven by the nucleus interface of

the nidopallium Abbrev. Title: NIf drives bursts in the sleeping songbird Authors:

Richard H.R. Hahnloser1, Michale S. Fee2

Addresses: 1: Institute of Neuroinformatics UZH / ETHZ, Winterthurerstrasse 190, 8057 Zurich, Switzerland. 2: Department of Brain and Cognitive Sciences and McGovern Institute for Brain Research,

Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139, USA.

Correspondence: Richard H.R. Hahnloser Institute of Neuroinformatics UZH/ETHZ Winterthurerstrasse 190 8057 Zurich Email: rich@ini.phys.ethz.ch +41 44 635 3060 phone +41 44 635 3053 fax Content: 2 Tables, 7 Figures, 31 pages Keywords:

zebra finch, sequence, efference copy, birdsong, neural code, antidromic

Acknowledgements: We thank Maria Minkoff and three anonymous reviewers for comments on the manuscript. Grants: These studies were supported by a Swiss National Science Foundation grant to R.H. and National Institute of Mental Health Grant RO1-MH067105 to M.S.F.

2

ABSTRACT

The function and the origin of replay of motor activity during sleep are

currently unknown. Spontaneous activity patterns in the nucleus robustus of the

arcopallium (RA) and in HVC of the sleeping songbird resemble premotor patterns

in these areas observed during singing. We test the hypothesis that the nucleus

interface of the nidopallium (NIf) has an important role for initiating and shaping

these sleep-related activity patterns.

In head-fixed, sleeping zebra finches we find that injections of the GABAA-

agonist muscimol into NIf lead to transient abolishment of premotor-like bursting

activity in HVC neurons. Using antidromic activation of NIf neurons by electrical

stimulation in HVC, we are able to distinguish a class of HVC-projecting NIf

neurons from a second class of NIf neurons. Paired extracellular recordings in NIf

and HVC show that NIf neurons provide a strong bursting drive to HVC. In

contrast to HVC neurons, whose bursting activity waxes and wanes in burst epochs,

individual NIf projection neurons are nearly continuously bursting and tend to

burst only once on the time scale of song syllables. Two types of HVC projection

neurons, premotor and striatal projecting, respond differently to the NIf drive, in

agreement with notions of HVC relaying premotor signals to RA and an

anticipatory copy thereof to areas of a basal ganglia pathway.

3

INTRODUCTION Studies of motor-related neural activity and their widely observed replay during

sleep are of a broad interest as they relate to the natural mechanisms for encoding and

consolidation of memories (Nadasdy et al., 1999; Fenn et al., 2003; Palchykova et al.,

2006; Walker and Stickgold, 2006). Premotor activity in the songbird nucleus robustus of

the arcopallium (RA) is correlated with song vocalizations on a temporal scale in the

submillisecond range (Yu and Margoliash, 1996; Chi and Margoliash, 2001; Leonardo

and Fee, 2005). RA neuron activity has also been recorded during sleep, where it strongly

resembles song-related activity (Dave et al., 1998; Dave and Margoliash, 2000). Song

and sleep-related spike patterns are believed to be generated by common neural

mechanisms, since during both singing and sleep, RA spike patterns are driven by sparse

sequences in HVC projection neurons (Hahnloser et al., 2002; Hahnloser et al., 2006).

Given that HVC projection neurons are of such importance to RA activity in the singing

and sleeping states, we speculate that projection neurons upstream of HVC are of similar

importance to the generation of sparse HVC sequences. Here we are interested in

elucidating the feedforward drive provided to HVC during sleep from the nucleus

interface of the nidopallium (NIf).

HVC receives input predominantly from the thalamic nucleus uveaformis (Uva)

in the thalamus, the medial magnocellular nucleus of the anterior nidopallium (MMAN),

and NIf(Nottebohm et al., 1982; Fortune and Margoliash, 1995; Vates et al., 1996; Rosen

and Mooney, 2006). NIf is embedded in the auditory field L complex and is believed to

be important for auditory-motor interactions in the song-control system. Lesions of NIf

have been found to abolish selectivity in HVC neurons to acoustic playback of the bird’s

own song (BOS) (Janata and Margoliash, 1999; Coleman and Mooney, 2004; Cardin et

al., 2005). Similarly, spontaneous activity in HVC of anesthetized birds is strongly

reduced after reversible or irreversible NIf lesions (Coleman and Mooney, 2004; Cardin

et al., 2005).

We are interested in elucidating the sleep-related activity patterns of NIf neurons

projecting to HVC (NIfHVC neurons). These NIf neurons were previously identified in

anesthetized birds and found to be more densely firing during BOS playback than HVC

4

projection neurons (Coleman and Mooney, 2004). Our goal is to identify potential

differences in the manner in which these NIf neurons drive activity patterns in

downstream areas in the premotor and the anterior forebrain pathways (Okuhata and

Saito, 1987) involved in song production and song learning (Bottjer et al., 1984; Sohrabji

et al., 1990; Scharff and Nottebohm, 1991; Doupe, 1993; Doupe et al., 2004). To

characterize the relationship of sleep-related activity patterns in higher premotor brain

areas, we study coherency functions from paired recordings of NIfHVC neurons and HVC

projection neurons, and study the effects of reversible pharmacological inactivation on

sleep-related spike trains. Our study reveals new insights into the dynamics of sleep-

related activity patterns and provides the first evidence of a target-specific population

code of NIf neurons. Parts of this paper have been published previously in abstract form

(Fee et al., 2002).

MATERIALS AND METHODS Our experiments were performed in head-fixed zebra finches, a preparation that

permits us to perform paired extracellular recordings from identified neurons and

pharmacological manipulations. General methods and procedures for the experimental

setup of head-fixed, sleeping birds and spike train data analysis have been described

previously (Hahnloser et al., 2006).

Subjects. Zebra finches (Taeniopygia guttata) were obtained from commercial

suppliers (Old Bridge, NJ (USA) and Animal Diffusion, Villarimboud (Switzerland)).

Animals were maintained on a day-night reversed 12 hour light cycle to assist in

obtaining sleep during daytime experimental sessions. Data were taken from a total of 36

adult zebra finches (>90d).

Surgery. Surgery began approximately at the onset of the night cycle. Birds were

anesthetized with 1-3% isoflurane in oxygen. A small hole (~200 μm) was made in the

dura over each area and the uncovered brain was protected with 2% low melting agarose

(Sigma, Inc). Bipolar stimulation electrodes (Teflon-insulated 50 μm diameter stainless

steel wire spaced 0.5 mm apart) were inserted into HVC and RA or Area X and secured

to the scull with dental acrylic. Wound margins were treated with lidocaine gel. The

animal was placed in a small foam restraint and subsequently moved to the recording

5

apparatus without further anesthesia. In some of the experiments, to promote sleep, the

bird was given a single dose (1-10 μg) of melatonin (Sigma) right after surgery, injected

subcutaneously in phosphate buffered saline.

Electrophysiology

Antidromic activation. Bipolar stimulation electrodes in HVC were used for

antidromic identification of NIf neuron type and RA (or Area X) stimulation electrodes

were used for identification of HVC neuron type (Fig. 1A). Electrical stimulation was

produced using an isolated stimulation unit (AMPI, Inc) with intensities varying from 50-

500 μA. For most experiments, single monophasic current pulses of 0.2 ms duration were

used as stimuli. Single or paired single-unit recordings were made in HVC and NIf. High

signal-to-noise (>10:1) recordings were made using sharp glass microelectrodes (5–15

MΩ, borosilicate, 1.0 mm OD, 0.7 mm ID) pulled on a vertical electrode puller (Model

730, Kopf Instruments, Inc) and filled with 2 M NaCl. Signals were amplified using a

Neurodata IR285 (Cygnus Instruments, Inc) or Axoclamp-2B (Molecular Devices, Inc)

intracellular amplifiers. Custom electronics were used for additional amplification (gain

of 100) and filtering (300 Hz high-pass 5-pole Bessel filter, 10 kHz low-pass constant

delay filter). Extracellular signals were digitized to 16 bits precision at a sampling rate of

20 kHz and stored in blocks of 200 seconds on a Pentium-based PC running custom

Labview software (National Instruments, Inc.).

NIf identification: For the NIf identification experiment in Figure 1B, we first

pressure injected 200 nl of 10% biotinylated dextran amines (BDA) in phosphate-

buffered saline into right HVC. The purpose of this tracer injection was to fluorescently

label NIf neurons. Two days later, under anesthesia, we penetrated NIf from a posterior

angle with a 1 3− MΩ platinum tungsten electrode. Every 20 μm between 2 and 3 mm

penetration depth, we recorded the responses to 30 HVC stimulations at 1 second

intervals (~ 200 μA amplitude). After assessing the extent of NIf by observing the

stimulation responses on an oscilloscope, we made two electrolytic lesions above and

below NIf, 900 μm apart. Lesions were made by injecting a constant current of 15 μA

for approximately 12 seconds. These lesions were later identified in histological sections

in order to precisely locate the electrode track within sagittal slices of the brain.

6

Stimulation response shown in Figure 1B was quantified by the root-mean squared

(RMS) Voltage amplitude in the 1 7− ms interval following stimulation onset (averaged

over the 30 stimulations per site).

NIf neuron identification: NIf neurons were found and isolated using ongoing 1

Hz stimulation in HVC to elicit spike responses. Once a cell was isolated, stimulation

threshold current was measured and stimulus intensity was set between 10% and 50%

above threshold to produce a reliable response. Continuous analog records were acquired

for 60 stimuli at 1 second intervals. The response latency of an antidromic spikes is

defined as the interval between stimulus onset and the first evoked spike. Latency

variability is defined as the standard deviation of response latency. We estimate that the

contribution to the latency variability due to our fixed sampling rate of 20 kHz is less

than 16 μs (dashed line in Fig. 1E). Because neurons in neighboring field L may also

project to HVC (Fortune and Margoliash, 1995), we tried to avoid recording from field-L

neurons by recording only from those NIf regions with very dense multiunit response

(hash) to stimulation, shown to correlate well with NIf extent (Fig 1B). Furthermore, we

observed that field L neurons projecting to HVC (LHVC neurons) had longer response

latencies to HVC stimulation than did NIfHVC neurons.. Also, LHVC neurons showed

strong phasic responses to brief sound clicks, whereas NIfHVC neurons did not. For the

NIf neurons classified as interneurons, we were similarly conservative and applied this

classification only to neurons found at penetration depths within the

electrophysiologically identified borders of NIf.

Reversible lesions in NIf. The effect of lidocaine and muscimol injections in NIf

on sleep-burst rate in HVC interneurons was examined by micro-injection of lidocaine or

muscimol in 2% phosphate-buffered saline during single-unit recordings of HVCI

neurons. Injections were made from pulled glass pipettes (50 μm tip size) using a

pressure injection system (Pico Spritzer, Inc.). Injected volumes were 10–100 nl. Control

injections were made in NIf with phosphate buffered saline. Typically, injection pipettes

were placed in the center of NIf after an electrophysiological recording session during

which the extent of NIf was well characterized by multiple electrode penetrations. To

measure the effect of the injection, we counted the number of bursts in the HVCI neuron

in the minute before and after the injection. Statistical significance of burst rate

7

differences before and after the injection was assessed using the Mann-Whitney U test

(p<0.001). In one case the bird woke up right after the injection (by observation of eye

opening). Because the burst rate in HVC neurons is known to be highly state dependent

(Hahnloser et al., 2006), data from this trial were discarded.

Histology: After each experiment, the animal was euthanized by

intramuscular injection of 20% urethane or nembutal. For the NIf identification

experiment, birds were transcardially perfused with 4% paraformaldehyde (PFA) in 25

mM NaPO4 buffer. The brain was removed for histological examination of Nissl-stained

(or unstained) slices to verify the location of stimulating and recording electrodes, and

drug injection sites. All experiments were carried out in accord with protocols approved

by the local IACUC and the Veterinary Amt of the Canton of Zurich, Switzerland.

Spike train Analysis.

All spike train analysis procedures were implemented with custom-made Matlab

(Matworks Inc., Natick MA, USA) scripts.

Instantaneous firing rate: One way of representing the activity at time t of a

neuron is by the instantaneous firing rate R(t), a continuous function defined by the

inverse of the interspike interval surrounding time t .

Burst rate: We identified a group of at least two spikes as a burst if the

instantaneous firing rate continuously exceeded 100 Hz. The burst rate is defined as the

number of events per unit time in which the instantaneous firing rate crosses the 100 Hz

barrier from below.

Interspike interval (ISI) probability density function (pdf): For each neuron A we

computed the ISI pdf ( )Ah τ as a basic characterization of its spike train (τ stands for the

ISI). First we computed the ISI histogram with bin centers iτ chosen on a logarithmic

scale ( 1,...,100)i = . The ISI pdf ( )A ih τ is simply a normalized ISI histogram satisfying

( ) 1.A ii

h τ =∑

8

Auto-covariance function: The auto-covariance function ( )AAC t of a spike

train ( )A tρ (modeled as a sum of delta functions) is a measure of spike density

fluctuations. It is computed as

2

0

1( ) ( ) ( )T

AA A A AC t s t s dsT t

ρ ρ ρ= + −− ∫ , (0.1)

where Aρ is the mean firing rate of neuron A, and T is the duration of the

recording (the denominator T t− results in an unbiased estimation of auto-covariance at

large times t T ). The first term on the right side of Equation (0.1) is known as the

auto-correlation function. By subtracting the term 2

Aρ in Equation (0.1), the asymptotic

value of the auto-covariance function is zero. For better visibility, we have plotted auto-

covariance functions on a logarithmic time scale it in Figures 3C and 3D by integration

between adjacent points, i.e., plotted is the discrete function 1( ) ( ) .i

i

t

AA AAtC i C t dt+= ∫

To assess the complexity of spike trains, we probed them for renewal dynamics,

i.e. we tested whether interspike intervals were generated by random sampling from the

ISI pdf ( ).Ah τ Renewal spike trains have no memory, as each ISI is a random variable

that is independent of the previous ISI. To perform this test, we have computed renewal

auto-covariance functions ( )RAAC t derived from the ISI pdfs under renewal assumptions

(here superscript ‘R’ stands for renewal), and compared these functions to the measured

auto-covariance functions. For a derivation of ( )RAAC t see e.g. (Gerstner and Kistler,

2002). Any deviation of ( )AAC t from ( )RAAC t is indicative of a memory in the spike train,

causing ISIs to be conditional on previous ISIs. A deviation at time it is significant if

( )RAAC i differs from ( )AAC i by more than three times the standard error of ( )AAC i (the

standard error was computed by jackknifing - see the Methods section on the coherency

analysis).

We also tested whether auto-covariance functions differed from renewal functions

on the population level. Average auto-covariance functions ( )AA AC t and ( )R

AA AC t for

9

the two NIf neuron types are shown by the full and dashed lines in Figures 3C and 3D. At

each time it we tested whether the auto-covariance and renewal functions at that time had

identical medians using the Wilcoxon rank sum test (p<0.01). Time intervals where

deviations were significant are indicated by thick horizontal bars in Figures 3C and D.

Coherency analysis: We have quantified correlations between spike trains in

simultaneously recorded cells by the coherency function ( )AB tγ , which is similar to the

cross-covariance function, but includes an additional normalization by the auto-

covariance functions of the two spike trains. This normalization serves to discount for

correlations arising from bursting tendencies of neurons (Thomson and Chave, 1991;

Kimpo et al., 2003). In the frequency domain ω , the coherency function ( )ABγ ω is

computed by normalizing the cross-covariance function ( )ABC ω between neurons A and

B by the square-roots of the auto-covariance functions:

( )( )( ) ( )

ABAB

AA BB

CC C

ωγ ωω ω

= . (0.2)

We determined the coherency in the time domain ( )AB tγ by inverse Fourier

transformation of ( ).ABγ ω As in our previous study (Hahnloser et al., 2006), we

smoothed the coherency function ( )AB tγ by convolution with a Gaussian windowing

function of 4 ms standard deviation. Finally, the smoothed coherency functions were

down-sampled to 1 ms temporal resolution by summing over coherencies in 1 ms bins.

Our finding do not critically depend on the parameter values for smoothing and down-

sampling.

To depict the typical coherency function between neuron types, we have plotted

the average coherency function ( ) ,AB ABtγ Figures 6C, 6E, 7B, and 7D.

Significance: The significance threshold for the coherency function of a

particular neuron pair was assessed by jackknifing the data in 10 second data windows

and computing the jackknife variance (Thomson and Chave, 1991)

( )22

1

1( ) ( ) ( )N

iJ AB AB

i

Nt t tN

σ γ γ=

−= −∑ ,

10

where N represents the number of data windows (depending on the duration of the

recording), ( )iAB tγ is the coherency function of the jackknifed data with the ith window

missing, and ( ) ( )iAB AB i

t tγ γ= is the average coherency function of the jackknifed data.

We set the significance threshold of measured coherencies to 3 ( )J tσ , shown by dashed

lines in Figures 6 and 7. This threshold corresponds to a roughly 99% confidence for the

significance test. Whenever the coherency function exceeded the significance threshold, a

peak emerged. The distributions of peak amplitudes versus peak latencies are shown in

the summary plots beneath the average coherency functions in Figures 6C, 6E, 7B, and

7E.

We have also analyzed coherency functions on a population level, to see whether

they had some characteristic shape. For all neuron pairs of given types, the average

coherency function was compared to the average significance threshold shown by dashed

lines in Figures 6C, 6E, 7B, and 7D. Characteristic peaks in average coherency were

found by testing whether differences between the coherency functions and the

significance thresholds had zero median using the Mann-Whitney signed rank test

(p<0.01). Time intervals where the median coherency function was significantly larger

than the median renewal function are indicated by thick horizontal bars in Figures 6C and

7D.

RESULTS We set out to characterize the role of NIf input for generating and shaping sleep-

related spontaneous activity in identified HVC neurons. As different types of NIf neurons

have not been previously classified, and their firing properties and correlations with

identified HVC neurons during sleep are unknown, we (1) distinguished NIfHVC neurons

from other NIf neuron types using antidromic activation by electrical stimulation in

HVC, (2) analyzed spike train statistics of identified NIf neurons, and (3) assessed spike

train correlations and coherencies of identified NIf and HVC neurons. Furthermore, to

characterize the properties of population activity in NIf and assess its relevance for

shaping spike trains in HVC, we (4) pharmacologically blocked NIf activity while

recording spike trains from identified HVC neurons. All our recordings and

11

manipulations were performed in the sleeping, head-fixed bird (see Methods). A

schematic of the zebra finch brain showing NIf and HVC and the location of stimulation

and recording electrodes is shown in Figure 1A.

NIf identification

NIf is surrounded by auditory areas of the field L complex (L1, L2, and L3).

Since neurons in field L1 and L3 have projection targets in the shelf of HVC, one could

imagine confounding them with NIf neurons when searching for antidromic responses to

HVC stimulation. However, there exist cytoarchitectural and morphological differences

between NIf and areas of the field L complex. Cytoarchitectural studies have revealed

increased cell density in NIf in comparison to L1 and L3, and morphological studies have

revealed differences in predominant Golgi cell types found in NIf and field L (Fortune

and Margoliash, 1992; 1995). Based on these studies we expected to find increased

density of extracellular responses to HVC stimulation when placing the recording

electrode inside NIf.

By quantifying the multiunit response density (stimulation hash) by the root-

mean-squared (RMS) voltage trace in a small time interval following HVC stimulation,

we found indeed a robust increase in RMS voltage when the recording electrode was

positioned within NIf (Fig. 1B). In this experiment, we verified the passage of the

electrode track through NIf by examining histological sections of the brain in which NIf

was retrogradely labeled by fluorescent indicator (see Methods). In the following

experiments, we used the clearly visible increase in stimulation hash at short stimulus

latencies as an operational assessment of NIf extent. All single unit recordings were

performed at penetration depths within this antidromically characterized region.

Classification of NIf neurons

We identified two distinct NIf response patterns to HVC stimulation (Fig. 1C, D).

Putative HVC-projecting NIf neurons responded to low-intensity stimulation (50-500

μA) in HVC with a single spike. Near-threshold stimulation produced latencies to the

first spike between 0.9 and 3.2 ms (average 1.6±0.6 ms, n=20) with small latency

variability in the range 14-69 μs (37±16 μs, n=20, Fig. 1E). Putative NIf interneurons

12

responded to low-intensity stimulation in HVC, with an increasing number of spikes for

increasing stimulus intensity. Near-threshold stimulation produced latencies to the first

spike ranging from 1.8 to 5.4 ms (mean 3.2±1.4 ms, n=5) and a large latency variability

in the range 236–887 μs (559±317 μs, n=5, Fig. 1E). On the basis of these results, we

infer that the neurons with small latency variability were HVC-projecting (NIfHVC)

neurons. For simplicity, we refer to the neurons with large latency variability as NIf

interneurons (NIfI neurons), although we cannot rule out that these neurons project

outside of NIf. In agreement with this classification, in all cases tested (n=20), we found

that neurons with latency variability less than 100 μs (putative NIfHVC neurons) exhibited

spike collisions when stimulated immediately after a spontaneous spike (Fig. 1C bottom).

Neurons with latency variabilities larger than 100 μs (putative NIfI neurons), on the other

hand, did not exhibit collisions (Fig. 1D bottom). The vast majority of neurons isolated

within the borders of NIf, as defined by the presence of antidromic hash, had small

latency variability and had firing patterns as described below for NIfHVC neurons. Long

latency variability neurons were rarely found, and all of these had firing patterns similar

to those described below for NIfI neurons. In the subsequent experiments, we identified

each recorded NIf cell by its antidromic stimulation response, distinguishing NIfI neurons

from NIfHVC neurons by their differences in spike latency variability.

The spike widths, measured at 25% of the peak amplitude, were smaller for NIfI

neurons (0.16 0.02± ms, range 0.13 0.19− ms, n=9) than they were for NIfHVC neurons

(0.20 0.04± ms, range 0.13 0.32− ms, n=51, Fig. 1F). However, based on the

considerable overlap between spike width distributions, we were unable to classify

neuron type based on spike width or even spike waveform (using principal component

analysis).

In many simultaneous recordings in NIf and HVC, we also identified HVC

neuron type by RA stimulation. In the course of these experiments, we found two neurons

in NIf (in two birds) that projected both to HVC and to RA, as judged by the spike

collision test. However, we were not able to determine whether these neurons formed

monosynaptic connections with both HVC and RA neurons; or whether the axonal

pathways of these HVC-projecting neurons simply passed in close proximity to RA, but

13

without targeting RA neurons. Note that RA has not been identified as a projection target

of NIf neurons in the literature.

Spike train properties of two NIf neuron types

During sleep, a large fraction of spikes in both NIf neuron types fell into high

frequency bursts (Fig. 2A,B). Nevertheless, NIfHVC and NIfI neurons differed

considerably in their spontaneous firing patterns during sleep. For example, the average

firing rate of NIfHVC neurons was 5.4 Hz, whereas that of NIfI neurons exceeded 27 Hz.

NIfHVC neurons fired fewer bursts per second and exhibited larger average instantaneous

firing rates during bursts than did NIfI neurons (Wilcoxon rank sum test, n=117 NIfHVC

neurons and n=13 NIfI neurons, p<0.01). Average firing rates, firing rates during bursts,

burst rates, burst widths, number of spikes per burst, and the fraction of total spikes that

are part of bursts are summarized in Table 1.

In our previous recordings we found that firing rates in HVC and RA neurons

waxed and waned, altering between a state of high firing lasting 1-2 seconds and a state

of low firing also lasting several seconds (Hahnloser et al., 2006). We previously

identified a suitable time scale to measure these firing-rate fluctuations to be 3 seconds.

When analyzed in similar 3 second windows, we did not find similarly large fluctuations

of firing rates in NIf neurons. Average firing rates of NIfHVC neurons were distributed in

the range from 0 Hz up to 22 Hz (filled circles in Fig. 2C). This range of firing rates is

comparable to that of HVCX neurons, a HVC neuron type of average firing rate less than

a third of that of NIfHVC neurons. NIfI neurons rarely displayed firing rates less than 5 Hz

or larger than 50 Hz (Fig. 2D). In comparison, HVCI neurons, despite having half the

average firing rate, achieved about twice the range of NIfI firing rates. In Figures 2C and

D, we also show firing rate distributions of NIf neurons in burst spike trains (spike trains

for which all single spikes not part of a burst have been removed for the analysis),

verifying that on a coarse time scale, small firing-rate fluctuations also applied to bursts.

This lack of extensive firing-rate fluctuations in NIf motivated us to measure

spike train statistics in NIfHVC neurons also in awake birds, as judged by their open eyes

after touching their tails (n=3 birds). In data from 3-minute recording sessions, we found

that firing rates and burst rates of NIfHVC neurons in the awake state were not statistically

14

different from the firing rates measured during sleep (Wilcoxon rank sum test, n=5

NIfHVC neurons, p<0.01; Table 1). Note that in our previous sleep studies in HVC and

RA, we never observed bursts in awake birds in any neuron type.

Interspike-interval (ISI) and auto-covariance functions

We decided to further study NIf spike trains and the biophysical mechanisms

responsible for their generation by examining interspike interval (ISI) probability density

functions (pdfs), and autocovariance functions. These functions are computed and

analyzed in the following.

The average ISI pdf of each NIf neuron type is bimodal, with a common first peak

corresponding to the ISIs during bursts and a second peak roughly corresponding to the

interburst interval (Fig. 3A, B). The average ISI pdf of NIfHVC neurons is reminiscent of

that of HVC projection neurons, as these neurons also emit virtually no spikes at an ISI of

about 10 ms (Hahnloser et al., 2006). Similarly, the average ISI pdf of NIfI neurons is

reminiscent of that of HVCI neurons.

In HVC projection neurons there is evidence of intrinsic (not synaptically driven)

bursting as judged by the stereotypy of recorded sleep bursts (Hahnloser et al., 2002),

whereas in HVC interneurons and RA neurons, there is no such evidence (Dave and

Margoliash, 2000). We investigated the burst stereotypy of NIf neurons by computing

auto-covariance functions. In 83 of 117 NIfHVC neurons, we found an oscillatory behavior

of auto-covariance functions on a very short time scale, with an example shown in the

inset of Figure 3C. This oscillatory behavior is a consequence of stereotyped burst

patterns and suggests a cellular rather than a network mechanism for burst generation in

NIfHVC neurons. Only 1 of 13 NIfI neurons showed an oscillatory auto-covariance

function, thus providing weaker evidence for intrinsic bursting mechanisms in NIfI

neurons (a typical non-oscillatory auto-covariance function is shown in the inset of Fig.

3D).

The simplest generative process of a spike train with a given ISI pdf is the

renewal process, according to which each ISI is drawn randomly and independently of

other ISIs from a fixed ISI pdf. By exploring whether NIf neuron spike trains are

compatible with a renewal process for their generation, we derived from their ISI pdfs

15

expected auto-covariance functions under renewal assumptions (see Methods). By

comparison to these renewal (auto-covariance) functions, we found further evidence for

intrinsic bursting tendencies in NIfHVC neurons. Of the 109 NIfHVC neurons tested, 57

exhibited auto-covariance functions that were smaller than the equivalent renewal

functions somewhere in the time interval 3-45 ms (jackknife, p<0.01). Other than these

cases, no significant differences with renewal statistics were observed. This absence of

long sequences of short interspike intervals suggests a limitation of NIfHVC neurons to

produce long bursts. In this respect, NIfHVC neurons are similar to HVC projection

neurons, in which a similar trend was seen.

A different behavior was seen in NIfI neurons. Of the 13 NIfI neurons tested, 10

exhibited auto-covariance functions compatible with renewal assumptions, and 3

exhibited auto-covariance functions larger than renewal functions somewhere in the

interval 3 31− ms. This behavior of NIfI neurons is reminiscent of RA and HVCI

neurons, in which bursts were also found to be longer than expected by renewal statistics.

Given these diverse behaviors of individual NIf neurons, we tested their

compatibility with renewal statistics also on the population level. We found that the

median auto-covariance function of NIfHVC neurons was smaller than the median renewal

function in the time interval 6 30− ms (Kruskal-Wallis test, p<0.01). However, the

median auto-covariance function of NIfI neurons was not significantly different from the

median renewal function (Kruskal-Wallis test, p<0.01, Fig. 3D). In conclusion, renewal

tests in two NIf neuron types reveal differences between their burst generation

mechanisms. Some form of depressive dynamics seems to shorten NIfHVC neuron bursts,

whereas no such mechanism appears to apply to NIfI neurons.

NIfHVC neurons display a firing decrement in their autocorrelation function. In the

range of about 10 100− ms, the average auto-correlation function of NIfHVC neurons

clearly dips below its asymptotic value (Fig. 3E). In this time range, NIfHVC neurons

exhibit a reduced spike probability down to 40% of baseline. We have not observed such

a firing decrement in any other neuron type in NIf, HVC, RA, or LMAN; it may

represent a soft refractory period unique to NIfHVC neurons.

16

Reversible lesions in NIf

To determine the role of NIfHVC population activity for the generation of sleep

bursts in HVC, we examined the effect of reversible inactivation of NIf while recording

activity in single HVCI neurons (Fig. 4). When we injected the sodium-channel blocker

lidocaine into NIf, the spontaneous bursting of HVCI neurons immediately stopped (Fig.

4A) and recovered after about 3-4 minutes (Fig. 4A, bottom). The HVCI neuron burst rate

in the 60s interval following the injection was significantly smaller than in the 60 second

interval preceding the injection (Mann-Whitney U test, p<0.01, n=3/3 birds, 3 injections

per bird, Fig 4C). Saline injections did not lead to a significant change in HVCI neuron

burst rates (Mann-Whitney U test, p<0.01, n=3/3 birds, 3 injections per bird). We

realized that a caveat of our lidocaine injections was that they might also have affected

fibers of passage from Uva to HVC. To restrict the effect of injections to neural

dendrites and somata in the NIf region, we also injected the GABAA agonist muscimol

into NIf. Similarly as for the lidocaine injections, we observed an immediate and

significant reduction in HVCI neuron burst rate in the minute following the muscimol

injection (Mann-Whitney U test, p<0.01, n=3/3 birds, 3 injections per bird, p<0.01, Fig.

4B). Recovery of burst rates was not observed within more than 20 minutes following

muscimol injections. Despite these reduced burst rates, we found that HVCI neurons kept

spiking after NIf inactivation (see Figures 4A and B), suggesting that lidocaine and

muscimol did not leak into HVC. Furthermore, the average firing rate of HVCI neurons in

the 50 seconds after lidocaine or muscimol injections into NIf was 10.2±5.4 Hz and so

was smaller than the typical firing of HVCI neurons during sleep: 15±9.0 Hz (Hahnloser

et al., 2006), but was not significantly different from average firing rates of these neurons

measured in the awake bird: 7.2±7.0 Hz (Hahnloser et al., 2006). These results suggest

that NIf provides essential input to HVC that is necessary to transform the regular spiking

of HVCI neurons in the awake state into bursting in the sleeping state. The reduced firing

and bursting of HVCI neurons after NIf inactivation agrees with previous inactivation and

norepinephrine injection experiments in the anesthetized bird (Cardin and Schmidt,

2004a; Coleman and Mooney, 2004).

17

Paired recordings in identified HVC and NIf neurons

To directly compare firing properties of NIf neurons with those of HVC neurons,

we performed paired recordings in these areas. NIf neuron type was identified by

antidromic stimulation in HVC, and HVC neuron type was identified by antidromic

stimulation in either RA or Area X.

From our single neuron recordings we know that NIfHVC neurons have a weak, but

significant tendency to form burst epochs (in Figure 3C, in the time interval 0.75 3.4−

seconds the median auto-covariance function is larger than the median renewal function,

Kruskal-Wallis test, p<0.01). However, as to be expected, in paired recordings of NIfHVC

and HVCI neurons, we observed much stronger burst epochs in HVCI neurons than in

NIfHVC neurons (Fig. 5A). To analyze co-fluctuations of firing rates in more detail, we

computed correlation coefficients of firing rates measured in non-overlapping 3 second

windows. For the neuron pair shown in Figure 5A, the correlation coefficient of 0.15 was

found to be significant ( 0.05).p < In fact, most NIfHVC–HVCI neuron pairs showed

significantly correlated firing rates (Fig. 5B, top left). By incorporating previously

recorded data in HVC into our analysis (Hahnloser et al., 2006), we found that NIfHVC-

HVCI firing-rate correlations are comparable to those of HVCRA-HVCI neurons (Fig. 5B,

bottom left). Similarly, firing rates in most NIfHVC–HVCRA neuron pairs were

significantly correlated, and also some NIfHVC–HVCX neuron pairs showed correlated

firing rates. Thus, on the level of firing-rate correlations with HVCI neurons, there is a

similarity between NIf and HVC projection neurons, despite their large differences in

burst rates.

The firing-rate correlations involving interneurons in NIf and HVC are stronger

and more reliable than those involving projection neurons in NIf and HVC. In 99 neuron

pairs not involving a projection neuron in either HVC or NIF, we only found one pair that

did not exhibit significantly correlated firing rates (Fig. 5B, right side). Average

correlation coefficients for all neuron pairs studied are summarized in Table 2.

To further test for similarities between NIfHVC and HVCRA neurons, we compared

firing rate distributions in paired recordings of NIfHVC and HVCRA neurons. Average

firing rates in these neurons differed by about one order of magnitude (NIfHVC: 5.4 3.3±

Hz, versus HVCRA: 0.6 .9± Hz, n=34 pairs), as did average burst rates (NIfHVC: 1.4 0.9±

18

Hz, versus HVCRA: 0.13 0.2± Hz, n=34 pairs). However, we found a striking similarity

between firing rate distributions of NIfHVC and HVCRA neurons after discounting for

baseline differences in firing rates. That is, we counted the number of spikes fired in 100

ms windows of burst spike trains by only taking into account windows where at least one

spike was found (thus discounting for differences in the number of empty windows). In

22 of 34 NIfHVC–HVCRA neuron pairs, we found indistinguishable firing rate distributions

in the two cells (Kolmogorov-Smirnov test, p<0.01). Similarly, the distribution of firing

rates of the entire NIfHVC population was identical to the distribution of the HVCRA

population (Kolmogorov-Smirnov test, p<0.01, Fig. 5C). Equality of firing rate

distribution in these different populations was robust and was observed for analysis

windows in the range of the soft refractory time of NIfHVC neurons, i.e., from 50 ms to

150ms. Firing similarity was less striking for other HVC neuron types. In 3 of 12 NIfHVC–

HVCX neuron pairs we found equal firing rate distributions (also not counting empty

windows). And, in only 3 of 37 NIfHVC–HVCI neuron pairs did we find equality of firing

rate distributions (likewise not counting empty windows). Equality of firing rate

distributions on the population level and in 100ms windows was only observed for

NIfHVC–HVCRA cell pairs.

To study the relationship between spikes in NIfHVC neurons and HVC projection

neurons on a high temporal resolution, we examined raster plots of NIfHVC neuron

activity aligned on HVCRA neuron bursts. We quantified the paired recordings by

coherency functions, derived from the full spike trains including single spikes. We found

that very often the HVCRA neuron bursts were preceded by NIfHVC neuron bursts by just a

few milliseconds (Fig. 6A, B). For most NIfHVC–HVCRA neuron pairs, we found one

significant peak in the coherency function (20/34 pairs exhibited each one significant

peak at median time lag of 5− ms, in the range 21− to 1 ms; 14/34 pairs exhibited no

significant peak, Fig 6C bottom). The average coherency function between NIfHVC and

HVCRA neurons exhibited a peak at 5− ms, well above the average significance threshold

(Fig. 6C, top). The median coherency function was above the median significance

threshold in the time interval from 7− to 5− ms (Wilcoxon signed-rank test, p<0.01).

Thus, NIfHVC neurons on average spike just a few milliseconds before HVCRA neurons

do.

19

We also examined raster plots of NIfHVC activity aligned on HVCX neuron bursts.

In many pairs, as was the case with HVCRA neurons, bursts in HVCX neurons tended to

be preceded by NIfHVC neuron bursts. However, in some cases, bursts in the HVCX

neuron preceded the NIfHVC neuron bursts (Fig. 6D). In summary, most NIfHVC–HVCX

neuron pairs exhibited at least one significant coherency peak (8/12 pairs exhibited at

least one significant peak at median time lag 2− ms, in the range 8− to 20 ms, Fig. 6E).

The three main peaks at a positive time lag in Figure 6E are from three different birds and

are unlikely to be a sampling artifact. The average coherency function between NIfHVC

and HVCX neurons exhibited two distinct peaks, one peak at 3.5− ms, well above the

average significance threshold, and the other at 15 ms, roughly coinciding with the

average significance threshold (Fig. 6E, top). From these data, differences between the

two types of HVC projection neurons are apparent. At a positive time lag larger than 5

ms there was no peak in 34 pairs with HVCRA neurons, whereas there were 6 peaks in 12

pairs with HVCX neurons. The only HVCRA neuron we found to somewhat reliably burst

before the NIfHVC neuron was the one shown in Figure 6B. However, as can be seen,

significance of the coherency function at a positive time lag was not achieved.

In paired recordings of NIfHVC neurons with HVCI neurons, we observed frequent

bursts in the HVCI neuron in close proximity to NIfHVC neuron bursts (Fig. 7A). In most

NIfHVC–HVCI neuron pairs there was at least one significant peak in the coherency

function (32/37 pairs exhibited at least one significant peak with a median time lag of 7

ms, range 24− to 123 ms, Fig. 7B bottom). In agreement with the notion that NIfHVC

neurons tend to burst before HVCI neurons, the average coherency function between

NIfHVC and HVCI neurons exhibited a peak at 5 ms, well above the average significance

threshold (Fig. 7B, top). The median coherency function of NIfHVC–HVCI pairs was

significantly larger than the median significance threshold in the time interval from 2 to

14 ms (Wilcoxon signed-rank test, p<0.01).

We were not able to record from many NIfI neurons. An example raster plot of a

NIfI neuron, simultaneously recorded with an HVCRA neuron is shown in Figure 7C.

There was a highly significant coherency peak as a result of the many NIfI neuron bursts

produced just a few milliseconds before the HVCRA neuron bursts. On the level of a small

20

population, we could verify this tendency (3/3 HVCRA–NIfI pairs exhibited at least one

significant peak with a median time lag of 9− ms).

For the paired recordings of NIfI neurons with HVCI neurons, we did not produce

raster plots due to the high burst densities in both of these neuron types. Coherency plots

peaked significantly at a small negative time lag with respect to HVCI neuron spikes (3/3

significant pairs with median time lag of 4− ms in the range 9− to 2− ms), showing that

the NIfI neurons were strongly coherent with HVCI neurons (Fig 7D).

In conclusion, the time lags of coherency peaks were in agreement with a causal

relationship between NIf neuron spikes and subsequent activity in HVC. Some of our

observations suggest interesting mechanisms by which NIf neurons drive bursts in HVC

neurons during sleep, discussed in the following section.

DISCUSSION We have found that NIf neurons can reliably be identified in extracellular

recordings by antidromic stimulation from HVC. Antidromic response latencies in

NIfHVC neurons were very short and narrowly clustered, suggesting that synchronized

bursts in NIf could lead to short latency bursts in HVC with very little temporal jitter.

Compared to NIfHVC neurons, NIfI neurons responded to HVC simulation with much

larger latency variability. We observed many distinguishing characteristics of

spontaneous activity in these two different NIf neuron types, including mean firing rates,

burst rates, and burst generation mechanisms.

Given the absence of any known projection target of NIf other than HVC, we

speculate that the second NIf neuron type belongs to a class of interneurons. One reason

for our speculation are the high burst rates in NIfI neurons, analogous to interneurons in

RA and HVC (Leonardo and Fee, 2005; Hahnloser et al., 2006). With RA- and HVC

interneurons being inhibitory (Spiro et al., 1999; Rosen and Mooney, 2003), it may well

be that NIfI neurons are inhibitory as well. This classification is quite plausible, given that

our GABA and muscimol injection experiments and previous such experiments (Cardin

and Schmidt, 2004b; Coleman and Mooney, 2004) strongly suggest an abundance of

GABA receptors in NIf. However, further experiments will be necessary to determine the

neurotransmitter type released by NIfI neurons.

21

We suspect the high coherencies between NIfI and HVC neurons arises from

common input from NIfHVC neurons, just as HVCI neurons are strongly coherent with RA

neurons, likely by virtue of common input from HVCRA neurons (Hahnloser et al., 2006).

Currently, the only evidence that NIfI neurons are synaptically driven by NIfHVC neurons

is that they respond with short latency to HVC stimulation, shown here to antidromically

activate NIfHVC neurons. More paired NIfI and HVC recordings would help to further

clarify the relationship between NIfI and HVC activity.

We did not find significant differences between firing rates of NIf neurons in the

awake and the sleeping states. This finding contrasts with a previous study where arousal

of lightly sedated birds lead to reduction of spontaneous NIf activity (Cardin and

Schmidt, 2004b). This apparent disagreement might arise from differences between the

sedated state and the sleeping state or might be attributed to effects of residual melatonin

in our experiments. In any case, we believe the most interesting phenomenon to be that

NIf cells burst in the awake state, a property that has never been observed in either RA or

HVC cells.

We speculate that there exists a mechanism that gates the response of HVC to

ongoing drive from NIf, much as there is a mechanism that gates the transmission of

auditory responses in HVC from NIf (Cardin and Schmidt, 2003; Cardin and Schmidt,

2004a).According to our view, NIfHVC neurons would constantly bombard and

occasionally activate HVC projection neurons. Not all of these events would necessarily

result in a dense sleep-burst epoch in HVC, to agree with the moderate firing-rate

correlations of NIfHVC and HVC projection neurons with HVCI neurons in Figure 5.

However, on some occasions, a dense sleep-burst epoch would form, activating most

HVC neurons. Since we found a weak but non-negligible tendency of NIfHVC neurons to

also form sleep-burst epochs (Figs. 3C, 5A, B), we speculate that Uva might somehow

control the generation of sleep-burst epochs, for example, by coordinating

neuromodulatory input (Akutagawa and Konishi, 2005). Further influences on HVC

excitability might also result from input from the medial magnocellular nucleus of the

anterior nidopallium (MMAN), as well as from direct cholinergic (Shea and Margoliash,

2003) or other neuromodulatory input to HVC (Appeltants et al., 2000).

To our surprise, despite large differences between average firing rates in NIf- and

22

HVC projection neurons, we have also found many similarities on a global and a local

level. First, NIfHVC and HVCRA neurons have a similar burst density on the level of the

entire population, when considering that burst rates in NIfHVC neurons were ten times

higher, but their population size is about ten times smaller. Also, thanks to their soft

refractoriness on the time scale of song syllables (~100 ms), firing-rate distributions of

NIfHVC neurons were found to be identical to those of HVCRA neurons (Fig. 5C).

A central question pertaining to NIf output is whether NIf activity serves only to

initiate a sequence of self-sustained sleep bursts in HVC or whether every single burst in

HVC is driven by a particular set of NIf bursts. In the former case, a particular set of

strongly synchronized bursts in NIf might serve as the single initialization signal for the

HVC bursts. In the latter case, transiently increased burst rates in NIf might lead to

repeated excitation of HVC neurons beyond their firing thresholds. Our paired recordings

involving NIfHVC neurons can be helpful to tease these two possibilities apart. Our main

observation was an interesting difference in the coherencies of NIfHVC neurons with the

two types of HVC projection neurons. HVCRA neurons burst reliably just after NIfHVC

neurons did, whereas HVCX neurons burst reliably after and before NIfHVC neurons. From

the coherencies with HVCRA neurons we would be led to conclude our first hypothesis —

that NIf activity serves to initialize HVC activity — is correct. However, the HVCX

bursts preceding the NIfHVC bursts suggest that, at least on the timescale of a few

milliseconds, NIf neurons engage in stereotyped sequential activity. Thus, the more

complete picture suggested by our data is that there is a first population burst in NIf that

activates mainly HVCX neurons. Then, a second population burst just a few milliseconds

later activates HVCRA and HVCX neurons. Such a scenario can explain why HVCRA

bursts are rarely observed to precede NIfHVC bursts, whereas HVCX neurons are observed

to frequently burst before and after NIfHVC neurons. It is interesting to note that, in our

previous recordings of HVC neuron pairs (Hahnloser et al., 2006), while most HVCX -

HVCRA pairs were not coherent, the single pair we found that showed coherent bursting

did so with the HVCX neuron leading by roughly 20 ms. Coherency differences between

the two types of HVC projection neurons might result from two distinct classes of NIfHVC

neurons, from HVC-intrinsic mechanisms such as inhibition (Mooney, 2000; Solis and

Perkel, 2005), or from other brain areas such as Uva. For example, Uva projection

23

neurons might directly drive HVCX neurons and indirectly drive HVCRA neurons via

activation of NIf.

We can speculate about the functional role of the observed latency differences

between HVCRA and HVCX neurons. It is clear that signals originating in HVC and

arriving in RA take longer if they do so by traveling along the anterior forebrain pathway

involving Area X (Okuhata and Saito, 1987) than by traveling directly from HVC to RA.

Thus, the anticipatory firing of HVCX neurons can partially compensate for signal delays

that would normally occur if they fired simultaneously with HVCRA neurons. The signal

propagation time from HVC to RA along the anterior forebrain pathway has been

estimated to be around 40 ms (Troyer and Doupe, 2000; Kimpo et al., 2003; Abarbanel et

al., 2004). With our measured time difference of 20 ms (15+5 ms) between the second

peak in HVCX coherency functions and the first peak in HVCRA coherency functions, it

follows that about half the signal propagation time along the anterior forebrain pathway is

compensated for by anticipated departure of HVCX neuron signals. With an estimated

propagation time of 4 ms for HVCRA signals to reach RA (Hahnloser et al., 2006), we

conclude that bursts in NIf tend to reach RA via the premotor pathway 16 ms before they

reach RA via the anterior forebrain pathway.

What could be the purpose of such a latency adjustment? Some previous models

of vocal learning have emphasized the importance of differences in propagation delay

through the premotor pathway and through the anterior forebrain pathway. In these

models HVCX activity is to be viewed as an efference copy of premotor signals in

HVCRA neurons (Troyer and Doupe, 2000; Abarbanel et al., 2004). Given the known

modulatory role of LMAN for song (Scharff and Nottebohm, 1991; Kao et al., 2005;

Olveczky et al., 2005), anticipation of HVC signals transmitted along the anterior

forebrain pathway could be important to maintain proper registration of premotor signals

along the two efferent pathways of HVC.

Given our acute experimental protocol, we were unable to relate individual spike

patterns in the sleeping bird to premotor activity during singing. Therefore, we analyzed

burst events, knowing that bursting in HVC and in RA is a signature of the singing state.

Our reversible inactivation experiments provide new insights into the role of NIf for burst

generation in HVC (Figure 4). Given the strong correlations between HVCI neurons and

24

other HVC neuron types during sleep (Hahnloser et al., 2006), and given the strong

reduction of auditory-evoked subthreshold activity in all HVC neuron types after NIf

inactivation (Coleman and Mooney, 2004), we speculate that NIf drives sleep bursts in

all HVC neuron types, and presumably also in RA. Such a role of NIf during sleep is

highly remarkable given that NIf neurons exhibit singing-related activity (McCasland,

1987; Hahnloser and Fee, 2003), but HVC seems to be able to produce song-related

bursts in adult zebra finches with bilateral NIf lesions (Cardin et al., 2005). In

combination with the loss of auditory responses in HVC neurons after NIf inactivation

(Coleman and Mooney, 2004), our results imply a surprising involvement of NIf in

conveying song-specific activity, both on a replay and on a sensory level. With the

possibility of existence of offline learning mechanisms during sleep (Nadasdy et al.,

1999; Margoliash, 2001; Deregnaucourt et al., 2005), we could imagine that auditory

hallucination of the BOS originating in NIf is the driving force of such memory

consolidation processes.

FIGURE CAPTIONS

Figure 1 NIf and NIf neuron identification. A) Schematics of the song-control system

and the experimental setup. Nucleus interface of the nidopallium (NIf) receives input

from the thalamic nucleus uveaformis (Uva) and projects to HVC. HVC projects to the

robust nucleus of the arcopallium (RA), which in turn innervates vocal motor neurons in

the hypoglossal nucleus (nXIIts). NIf neuron type was identified by antidromic

stimulation in HVC, and HVC neuron type by antidromic stimulation in RA or Area X.

B) NIf extent, as assessed by the region of increased response to stimulation agrees with

histological identification. NIf has been retrogradely labeled by BDA injection in HVC.

The two white circles close to the extremes of the white line mark electrolytic lesions

used as reference points along the electrode track. Shown in the tilted axes is the root-

mean square (RMS) extracellular voltage amplitude in the interval 1 7− ms following

stimulation onset. The RMS amplitude is high (>~1 mV) when the electrode is inside NIf

and low otherwise. C) Electrical stimulation in HVC elicits spike responses in NIfHVC

25

neurons at a single precise time lag (top). Stimulating 0.5 ms after a spontaneous spike in

a NIfHVC neuron leads to response failure (bottom). D) Antidromic responses in NIfI

neurons have larger latency variability (top) and do not fail when HVC is stimulated right

after a spontaneous spike (bottom). E) A plot of spike latency versus latency variability

allows for segregation of NIfHVC neurons from NIfI neurons based on a latency-variability

threshold of 0.1 ms. F) Spike waveforms of NIfHVC neurons (top) and of NIfI neurons

(bottom) are very similar.

Figure 2 NIf neuron spike trains during sleep. A) Extracellular records of a NIfHVC

neuron (top) and a NIfI neuron (bottom). The burst events labeled on the left are redrawn

on a smaller time scale on the right. B) Instantaneous firing rate (IFR) functions of a

NIfHVC neuron (top) and a NIfI neuron (bottom). C) Firing rate distribution of NIfHVC

neurons and D) NIfI neurons in three second windows. Filled symbols are average firing

rates in regular spike trains, and open symbols are from burst spike trains with single

spikes removed. The numbers shown on the right are average firing rates plus/minus

standard deviations for the burst spike trains, computed after removing single spikes.

Figure 3 Interspike interval (ISI) probability density functions (pdfs) and autocorrelation

functions. A) Average ISI pdf of NIfHVC neurons, with the shaded area representing

plus/minus the standard deviation. There are few ISIs of ~10 ms and few ISIs larger than

one second. B) The average ISI pdf of NIfI neurons shows virtually no ISIs larger than

400 ms. C) Average auto-covariance function of NIfHVC neurons plotted on a logarithmic

time scale. The delta-function peak at time zero has been suppressed. The dashed line is

the average renewal auto-covariance function derived from the respective ISI pdfs. In the

interval close to 10 ms, the auto-covariance function is smaller than the renewal function,

and in the interval around 1 second, it is larger (as indicated by the two thick horizontal

bars below the auto-covariance function). The inset shows the auto-covariance function

of the NIfHVC neuron from Figure 2B, reflecting stereotyped bursting behavior by the

oscillatory shape. D) The average auto-covariance function of NIfI neurons is larger than

the renewal function, but this difference is not significant. The inset depicts the auto-

26

covariance function of the NIfI neuron from Figure 2B. E) Average autocorrelation

function of NIfHVC neurons plotted on a linear time scale, clearly displaying the soft

refractory period of about 125 ms during which the spiking probability is reduced down

to at most 40 percent of baseline.

Figure 4 Effect of lidocaine and muscimol injections into NIf on sleep burst rate in HVCI

neurons. The numbers separated by a backlash indicate the number of bursts measured in

the 50 second interval before the injection and after the injection. A) Lidocaine injections

into NIf immediately disrupt bursting in HVCI neurons as can be seen from the

instantaneous firing rates of two HVCI neurons (top and bottom). Bursting recovers

within 3-4 minutes (bottom). B) Muscimol injections into NIf have a similar effect of

reducing burst rate in HVCI neurons, but no recovery is observed on this timescale. C)

Median post-injection burst rates of HVCI neurons were significantly lower than median

pre-injection rates for both muscimol and lidocaine injections. Burst rates recovered after

5 minutes of lidocaine injections to values similar to pre-injection rates. The error bars

delimit the upper and lower quartiles and the asterisks denote significant differences

(Mann-Whitney U test, p<0.01).

Figure 5 Burst density variations in NIfHVC and HVCI neurons are weakly correlated. A)

Instantaneous firing rates in a simultaneously recorded NIfHVC–HVCI neuron pair. The

clustering of HVCI neuron spikes into burst epochs is only weakly reflected in the NIfHVC

neuron (two burst epochs in the HVCI neuron that seem to be coincident with increased

bursting in the NIfHVC neuron are marked by shaded areas). B) Correlation coefficients

of firing rates measured in 3 second windows of simultaneously recorded neuron pairs.

Black bars represent significantly correlated pairs (p<0.01). Correlations among neuron

pairs involving HVC or NIf projection neurons are often not significant and smaller (left

row) than correlations among other neuron pairs (right row). The arrow indicates the

correlation coefficient of the neuron pair in A. C) By performing the same analysis as in

Figure 2C, but on a time window of 100 ms instead of 3 seconds, we find identical firing-

rate distributions in NIfHVC and HVCRA neurons. To produce this subfigure we analyzed

burst-train data from 34 simultaneously recorded neuron pairs. As illustration, the

27

‘probabilities’ of zero spikes in NIfHVC and HVCRA neurons are also shown; these differ

strongly, but are not taken into account for comparing the distributions.

Figure 6 Paired recordings of NIf neurons and HVC projection neurons. A) Spike raster

plot of simultaneously recorded NIfHVC and HVCRA neuron pair. The spike bursts of the

HVCRA neuron (short red lines) are aligned on the first spike at the center of the plot.

Corresponding NIfHVC spikes (short black lines) are shown below each HVCRA burst.

NIfHVC neuron bursts tend to precede the HVCRA neuron bursts by a few milliseconds.

Below the raster plot is the coherency function plotted as a black line, with the

significance threshold shown by the dashed line. The full line exceeds the dashed line in a

peak at 8− ms, indicating significant coherency. B) Same analysis for another NIfHVC–

HVCRA neuron pair. Again, significant coherency exists at a negative time lag. The few

bursts in the NIfHVC neuron that follow the HVCRA neuron bursts are not significant as

can be seen in the bottom. C) Summary of all HVCRA–NIfHVC neuron pairs. Top: The full

line shows the average coherency function between all HVCRA–NIfHVC neuron pairs. The

shaded area delimits the standard deviation of coherency functions. The dashed line

shows the average significance threshold. There is a narrow coherency peak at a negative

time lag where the median coherency is larger than the median significance threshold

(thick horizontal bar). Bottom: All significant coherency peaks from individual neuron

pairs. D) Raster plot of a HVCX–NIfHVC neuron pair. The NIfHVC neuron bursts reliably

after the HVCX neuron (with significant coherency). E) Summary of HVCX–NIfHVC

neuron pairs. The average coherency function (top) exhibits significant peaks at 3.5− and

at 15 ms. Peaks of individual coherency functions are found at both positive and negative

time lags in the range 40− to 100 ms. Full circles denote the main peaks of individual

neuron pairs and open circles denote significant secondary peaks (of smaller coherency

value than the main peak).

Figure 7 Raster plots involving HVCI and NIfI neurons. A) Raster plot of HVCI neuron

spikes, time-aligned on NIfHVC neuron bursts. Below, the coherency function peaks 1 ms

after NIfHVC neuron spikes. B) The average coherency function of NIfHVC–HVCI neurons

28

peaks 5 ms after NIfHVC neuron spikes (top). Individual peaks are almost exclusively

located at positive time lags of NIfHVC neuron spikes. C) Raster plot of NIfI neuron

spikes, time-aligned to HVCRA neuron bursts. The coherency function significantly peaks

9− ms before HVCRA neuron spikes (bottom). D) The average coherency function of

HVCI–NIfI neuron pairs peaks 6− ms before HVCI neuron spikes (top). All three neuron

pairs exhibit significant coherency peaks at a small negative time lag (bottom).

TABLES

Neuron type Firing

rate (s-1)

Burst firing

rate (s-1)

Burst

rate (s-1)

Burst

width (ms)

Spikes per

burst

Spikes in

bursts (%)

NIfHVC (n=117) 5.4±3.3 461±139 1.4±0.9 4.4±1.7 2.7±0.5 70±20

Awake NIfHVC

(n=5)

5.0±1.9 550±173 1.2±0.6 3.7±1.2 2.5±0.3 58±19

NIfI (n=13) 27.8±9.6 321±87 4.9±2.9 7.5±1.7 3.0±0.9 53±24

TABLE 1: Firing statistics for two NIf neuron types. All values are means

plus/minus standard deviation.

Neuron pair # correlated pairs Corr. Coeff. Med. burst ratio

NIfHVC–HVCI 22/37, 13 birds 0.42 0.17± 1.2

HVCI–NIfI 3/3, 3 birds 0.51 0.15± 3.2

HVCRA–NIfHVC 19/33, 12 birds 0.28 0.11± 19.6

HVCRA–HVCI 16/26, 6 birds 0.42 0.12± 46.8

HVCX–NIfHVC 2/12, 4 birds 0.29 0.01± -

HVCI–HVCI 19/19, 5 birds 0.91 0.06± −

RA–RA 28/29, 3 birds 0.79 0.1± −

RA–HVCI 50/50, 9 birds 0.66 0.19± 3.4

29

TABLE 2: Sleep-related firing-rate correlations and burst ratios in different

neuron pairs, measured in non-overlapping 3 second windows. The criterion for

significance was 0.05.p < Correlation coefficients (mean plus/minus standard deviation)

were computed for significantly correlated pairs only. The median burst-rate ratios in the

last column were computed by the median value of ratios of average firing rates in

simultaneously recorded pairs of neurons, single spikes removed. In the denominator for

this ratio, we always choose the neuron appearing on the left in the first column. For

example, NIfI neurons fired 3.2 times as many spikes in burst trains as did HVCI neurons.

REFERENCES Abarbanel HDI, Gibb L, Mindlin GB, Rabinovich MI, Talathi S (2004) Spike timing and

synaptic plasticity in the premotor pathway of birdsong. Biol Cybern 91:159-167. Akutagawa E, Konishi M (2005) Connections of thalamic modulatory centers to the

vocal control system of the zebra finch. Proc Natl Acad Sci U S A 102:14086-14091.

Appeltants D, Absil P, Balthazart J, Ball GF (2000) Identification of the origin of catecholaminergic inputs to HVc in canaries by retrograde tract tracing combined with tyrosine hydroxylase immunocytochemistry. J Chem Neuroanat 18:117-133.

Bottjer SW, Miesner EA, Arnold AP (1984) Forebrain lesions disrupt development but not maintenance of song in passerine birds. Science 224:901-903.

Cardin JA, Schmidt MF (2003) Song system auditory responses are stable and highly tuned during sedation, rapidly modulated and unselective during wakefulness, and suppressed by arousal. J Neurophysiol 90:2884-2899.

Cardin JA, Schmidt MF (2004a) Noradrenergic inputs mediate state dependence of auditory responses in the avian song system. J Neuroscience 24:7745-7753.

Cardin JA, Schmidt MF (2004b) Auditory responses in multiple sensorimotor song system nuclei are co-modulated by behavioral state. J Neurophysiol 91:2148-2163.

Cardin JA, Raksin JN, Schmidt MF (2005) Sensorimotor nucleus NIf is necessary for auditory processing but not vocal motor output in the avian song system. J Neurophysiol 93:2157-2166.

Chi Z, Margoliash D (2001) Temporal precision and temporal drift in brain and behavior of zebra finch song. Neuron 32:899-910.

Coleman MJ, Mooney R (2004) Synaptic transformations underlying highly selective auditory representations of learned birdsong. J Neurosci 24:7251-7265.

Dave AS, Margoliash D (2000) Song replay during sleep and computational rules for sensorimotor vocal learning. Science 290:812-816.

Dave AS, Yu AC, Margoliash D (1998) Behavioral state modulation of auditory activity in a vocal motor system. Science 282:2250-2254.

30

Deregnaucourt S, P.P. M, Feher O, Pythe C, Tchernichovski O (2005) How sleep affects the developmental learning of bird song. Nature 433:710-716.

Doupe AJ (1993) A neural circuit specialized for vocal learning. Curr Opin Neurobiol 116:104-111.

Doupe AJ, Solis MM, Kimpo R, Boettiger CA (2004) Cellular, circuit, and synaptic mechanisms in song learning. Ann N Y Acad Sci 1016:495-523.

Fee MS, Hahnloser HR, Kozhevnikov AA (2002) The origin of a representation of time in the brain. In: Society for Neuroscience: Washington, DC.

Fenn KM, Nusbaum HC, Margoliash D (2003) Consolidation during sleep of perceptual learning of spoken language. Nature 425:614-616.

Fortune ES, Margoliash D (1995) Parallel pathways and convergence onto HVc and adjacent neostriatum of adult zebra finches (Taeniopygia guttata). J Comp Neurol 360:413-441.

Gerstner W, Kistler W (2002) Spiking Neuron Models. Single Neurons, Populations, Plasticity. Cambridge: Cambridge University Press.

Hahnloser HR, Fee MS (2003) Single neuron recordings in nucleus interface of singing zebra finches. In: Society for Neuroscience. Washington, DC: Abstract Vierwer/Itinerary Planner, Program 294.4.

Hahnloser HR, Kozhevnikov AA, Fee MS (2006) Sleep-related neural activity in a premotor and a basal-ganglia pathway of the songbird. J Neurophysiol 10.1152/jn.01064.2005.

Hahnloser R, Kozhevnikov A, Fee MS (2002) An ultrasparse code underlies the generation of neural sequences in a songbird. Nature 419:65-70.

Janata P, Margoliash D (1999) Gradual emergence of song selectivity in sensorimotor structures of the male zebra finch song system. J Neurosci 19:5108-5118.

Kao MH, Doupe AJ, Brainard MS (2005) Contributions of an avian basal ganglia-forebrain circuit to real-time modulation of song. Nature 433:638-643.

Kimpo RR, Theunissen FE, Doupe AJ (2003) Propagation of correlated activity through multiple stages of a neural circuit. J Neurosci 23:5750-5761.

Leonardo A, Fee MS (2005) Ensemble coding of vocal control in birdsong. J Neuroscience 19:652-661.

Margoliash D (2001) Do sleeping birds sing? Population coding and learning in the bird song system. Prog Brain Res 130:319-331.

McCasland JS (1987) Neuronal control of bird song production. J Neurosci 7:23-39. Mooney R (2000) Different Subthreshold Mechanisms Underlie Song Selectivity in

Identified HVc Neurons of the Zebra Finch. J Neurosci 201:5420-5436. Nadasdy Z, Hirase H, Czurko A, Csicsvari J, Buzsaki G (1999) Replay and time

compression of recurring spike sequences in the hippocampus. Journal of Neuroscience 19:9497-9507.

Nottebohm F, Kelley DB, Paton JA (1982) Connections of vocal control nuclei in the canary telencephalon. J Comp Neurol 207:344-357.

Okuhata S, Saito N (1987) Synaptic connections of thalamo-cerebral vocal nuclei of the canary. Brain Res Bull 18:35-44.

Olveczky BP, Andalman AS, Fee MS (2005) Vocal experimentation in the juvenile songbird requires a basal ganglia circuit. PLoS Biol 3:e153.

31

Palchykova S, Winsky-Sommerer R, Meerlo P, Durr R, Tobler I (2006) Sleep deprivation impairs object recognition in mice. Neurobiol Learn Mem.

Rosen MJ, Mooney R (2003) Inhibitory and excitatory mechanisms underlying auditory responses to learned vocalizations in the songbird nucleus HVC. Neuron 39:177-194.

Rosen MJ, Mooney R (2006) Synaptic interactions underlying song-selectivity in the avian nucleus HVC revealed by dual intracellular recordings. J Neurophysiol 95:1158-1175.

Scharff C, Nottebohm F (1991) A comparative study of the behavioral deficits following lesions of various parts of the zebra finch song system: implications for vocal learning. J Neurosci 11:2896-2913.

Shea SD, Margoliash D (2003) Basal forebrain cholinergic modulation of auditory activity in the zebra finch song system. Neuron 40:1213-1226.

Sohrabji F, Nordeen EJ, Nordeen KW (1990) Selective impairment of song learning following lesions of a forebrain nucleus in the juvenile zebra finch. Behav Neural Biol 53:51-63.

Solis MM, Perkel DJ (2005) Rhythmic activity in a forebrain vocal control nucleus in vitro. J Neurosci 25:2811-2822.

Spiro JE, Dalva MB, Mooney R (1999) Long-range inhibition within the zebra finch song nucleus RA can coordinate the firing of multiple projection neurons. J Neurophysiol 81:3007-3020.

Thomson DJ, Chave AD, eds (1991) Jackknifed error estimates for spectra, coherences, and transfer functions. Englewood Cliffs, NJ: Prentice Hall.

Troyer TW, Doupe AJ (2000) An associational model of birdsong sensorimotor learning. I. Efference copy and the learning of song syllables. J Neurophysiol 84:1204-1223.

Vates GE, Broome BM, Mello CV, Nottebohm F (1996) Auditory pathways of caudal telencephalon and their relation to the song system of adult male zebra finches. J Comp Neurol 366:613-642.

Walker MP, Stickgold R (2006) Sleep, memory, and plasticity. Annu Rev Psychol 57:139-166.

Yu AC, Margoliash D (1996) Temporal hierarchical control of singing in birds. Science 273:1871-1875.

A B

Figure 1

HVC

RA

nXIIts

Syrinx

Stim

Uva

NIf

Stim

Stim

Area XArea X

1ms 1ms

C D

Free

Spike-triggered

E

0 2 4 610

-2

10-1

100

Latency (ms)

Late

ncy

variabili

ty(m

s)

0.5ms

F

6ms

6ms

400

600

800

200

0

1

2

3

100 m�

NIfI

NIfI NIf

INIf HVC

NIf HVCNIf HVC

10s300Hz

B

Figure 2

C D

1 32

4

400ms 4ms

1 2 3

4

A

102

NIf HVC

#3

sw

ind

ow

s

3.7 � 2.5 Hz

100

NIf I

Firing rate (Hz)

13.8 �11.2 Hz10

1

0 50 10010

0

102

HVC I 8.11�7.22 Hz

0 20 40 6010

0

HVCRA

#3

sw

ind

ow

s

0.32�0.41 Hz

Firing rate (Hz)

102

NIfI

NIfI

NIf HVC

NIf HVC

100

E

A

Figure 3

0 50 125 2500

0.4

1

Time (ms)

Au

toco

rre

latio

n(b

ase

line

no

rma

lize

d)

B

C

5ms

5ms

100

101

102

103

104

-2

0

2

4

6

8x 10

-3

Time (ms)

Au

toco

va

ria

nce

100

101

102

103

104

-1

0

1

2

3

4x 10

-3

Time (ms)

Au

toco

va

ria

nce

100

101

102

103

104

0

0.05

0.1

Interspike Interval (ms)

Pro

ba

bili

ty

100

101

102

103

104

0

0.02

0.04

0.06

0.08

Interspike Interval (ms)

Pro

ba

bili

ty

DNIf

I

NIfI

NIf HVCNIf HVC

NIf HVC

0 20 40 60 80 1000

500

1000

Time (sec)

Firin

gra

te(H

z)

0 20 40 60 80 100

Time (sec)

Firin

gra

te(H

z)

500

1000

0 100 200 300 4000

500

1000

Time (sec)

Firin

gra

te(H

z)

0 20 40 60 80 100

Time (sec)

Firin

gra

te(H

z)

0

500

1000

A B

Figure 4

Nif injection

Lidocaine 20/1

Lidocaine 36/0

Nif injection

Nif injection

Muscimol 21/1

Muscimol 23/0

Nif injection

Pre

Pos

t

Rec

over

y

C **

0

0.3

0.6

Burs

tra

te(H

z)

10 s

0

5

10

#p

airs HVCI - HVCI

0

5

10RA - HVC

#p

airs

-1 -0.5 0 0.5 1Correlation coefficient

0

5

10

RA - RA

#p

airs

Figure 5

A

B

0

1HVCI - NIfI

#p

airs

0

5NIfHVC - HVCI

#p

airs

0 5 10

10-4

10-3

10-2

10-1

100

Pro

ba

bili

ty

CCorrelation coefficient

-1 -0.5 0 0.5 1

0

5

10HVCRA- NIf HVC

#p

airs

0

2

4HVC RA- HVCI

#p

airs

0

2

4HVCX - NIf HVC

#p

airs

I

NIf HVC

HVCRA

# spikes in 100 ms window

0HV

Cfiring

rate

(Hz)

500

0NIf

f.r.

(Hz)

500

IH

VC

NIf HVC

HVCRAA B

D

Figure 6

HVC RANIfHVC

HVCXNIf

HVC

C

E

-400 -200 0 200 4000

0.01

0.02

0.03

Time (ms)

Cohere

ncy

-400 -200 0 200 400

0

2

4

6x 10

-3

Time (ms)

Cohere

ncy

-400 -200 0 200 4000

0.01

0.02

Time (ms)

Cohere

ncy

-300 -200 -100 0 100 200 300

0

5

10

15

x 10-3

Time (ms)

Ave

rag

eco

he

ren

cy

HVCRA

- NIfHVC

-300 -200 -100 0 100 200 3000

0.01

0.02

Time (ms)

Peak-300 -200 -100 0 100 200 300

0

0.01

0.02

Time (ms)

Peak -300 -200 -100 0 100 200 300

0

5

10

x 10-3

Time (ms)

Ave

rag

eco

he

ren

cy

HVCX - NIfHVC

HVCRA

Figure 7

-400 -200 0 200 400

0

0.01

0.02

Time (ms)

Cohere

ncy

A C

-400 -200 0 200 400

0

5

10x 10

Time (ms)

Cohere

ncy

-3

NIfHVCHVCI

-300 -200 -100 0 100 200 300

0

0.01

0.02

0.03

Time (ms)

Ave

rag

eco

he

ren

cy

HVCI - NIf I

-300 -200 -100 0 100 200 300

0.01

0.02

0.03

Time (ms)

Peak

B

-300 -200 -100 0 100 200 300

0

5

10

15

x 10-3

Time (ms)

Ave

rag

eco

he

ren

cy

NIfHVC - HVCID

-300 -200 -100 0 100 200 300

0.005

0.015

0.025

Time (ms)

Peak

NIfI